Re: Lay common foundation to make PVR/SGX work without hacks on OMAP34xx, OMAP36xx, AM335x and potentially OMAP4, OMAP5

[Tero Kristo]

On 07/10/2019 18:52, Tony Lindgren wrote:
> Hi,
>
> * H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> [191005 16:59]:
> > Hi all,
> > with the arrival of v5.4-rc1 some of Tony's sysc patches have arrived
> > upstream, so we do no longer need them here.
> >
> > Therefore, I have rebased my drivers/staging/pvr driver [1] and fixed some
> > more issues:
> > * omap4 build only needs to distinguish between omap4420/30/60 and omap4470,
> >    because the latter has an sgx544 inside and the other sgx540
> >    This is solved by creating a new omap4470.dts
> > * I have added proper reg values and interrupts to the omap4 device
> >    tree node of the sgx (child node of the target-module)
> > * some updates to my sgxdump and sgxdemo scripts (assuming simple
> >    Debian Stretch rootfs)
> > * James Hilliard has contributed a fix for osfunc.c
> > * omap2plus also needs to be configured for STAGING and PREEMPT
> >    to be able to compile the driver
> > * I have added the __always_inline fix [2] which is needed for v5.4 with
> >    CONFIG_CC_OPTIMIZE_FOR_SIZE=y (which I are enabled on the Letux builds)
> >
> > Unfortunately Tero's rstctrl patches did not yet make it upstream (or even
> > linux-next) so I also have a copy in this branch.
> >
> > Results of first testing are:
> > * OMAP3530 (OpenPandora, BeagleBoard C): fails with
> > [  559.247558] PVR_K:(Error): SysLocateDevices: platform_get_resource failed
> >
> > * DM3730 (GTA04, BeagleBoard XM): kernel module loads
> >
> > * OMAP4460 (Pandaboard ES): kernel module loads
> >
> > * AM335x (BeagleBoneBlack): reports a problem with omap_reset_deassert:
> > [  204.246706] omap_reset_deassert: timedout waiting for gfx:0
>
> Please try with Tero's current github branch at github.com/t-kristo/linux-pm.git
> 5.4-rc1-ipc from few days ago, the earlier versions had still issues.

Yeah, this one should be fixed now.

> > * OMAP5 (Pyra): fails to enable the clocks (did work with the previous version)
> > [  304.140363] clock-controller:clk:0000:0: failed to enable
> > [  304.147388] PVR_K:(Error): EnableSGXClocks: pm_runtime_get_sync failed (16)
>
> Hmm no idea what might be up with this one. Did some clkctrl clock
> fixes maybe cause a regression here? Tero do you have any ideas?

So, this one I am not too sure, I haven't looked at omap5 graphics 
clocking. I don't think it has anything to do with reset handling though.

Is there some simple way to try this out on board; without PVR module 
that is?

-Tero

> > * OMAP5 with omap2plus_defconfig:
> > root@letux:~# echo on > $(find /sys -name control | grep \/5600)
> > [  213.490926] clock-controller:clk:0000:0: failed to enable
> > root@letux:~#
> >
> > * pvrsrvctl --start --no-module:
> >    reports (where the kernel module loads) that the uKernel does not run
> >
> > So I have several ideas what the reasons for the problems on the non-omap5
> > devices could be:
> > * initial code may have some omap5 specific hack inside
> > * or has omap5 specific magic constants
> > * uKernel may "know" on which platform it runs and
> >    we would need differently patched user-space code
> >    for each one
> > * omap5 has a dual core sgx544 while the other
> >    have single core
> > * the register address translation is not yet correct and
> >    this inhibits communicating of the user-space libs
> >    with the uKernel
> >
> > Maybe, if someone can point me to a complete and working BeagleBone source
> > tree (any kernel release) which makes use of 1.14.3699939 SDK, I could compare
> > code and address setup to find what makes the difference.
>
> Regards,
>
> Tony
>
> > [1]: https://github.com/openpvrsgx-devgroup/linux_openpvrsgx/commits/letux-pvr
> > [2]: https://lkml.org/lkml/2019/10/2/201

--
Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki. Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki